University of Pennsylvania
ScholarlyCommons
PSC Working Paper Series
Population Studies Center
12-9-2013
Schooling has Smaller or Insignificant Effects on
Adult Health in the US than Suggested by CrossSectional Associations: New Estimates Using
Relatively Large Samples of Identical Twins
Vikesh Amin
Central Michigan University
Jere R. Behrman
University of Pennsylvania, [email protected]
Hans-Peter Kohler
University of Pennsylvania, [email protected]
Amin, Vikesh, Jere R. Behrman, and Hans-Peter Kohler. 2013. "Schooling has smaller or insignificant effects on adult health in the US than suggested
by cross-sectional associations: New estimates using relatively large samples of identical twins." Philadelphia, PA: Population Studies Center, University
of Pennsylvania. PSC Working Paper Series, PSC 13-12. http://repository.upenn.edu/psc_working_papers/52
This paper is posted at ScholarlyCommons. http://repository.upenn.edu/psc_working_papers/52
For more information, please contact [email protected].
Schooling has Smaller or Insignificant Effects on Adult Health in the US
than Suggested by Cross-Sectional Associations: New Estimates Using
Relatively Large Samples of Identical Twins
Abstract
Adult health outcomes and health behaviors generally are strongly associated with schooling attainment. But
such associations do not necessarily imply that schooling has causal effects on health outcomes and behaviors
of the magnitudes of the associations. Schooling may be proxying for unobserved factors that are related to
genetics and family background. Recently several studies have used within-identical (monozygotic, MZ)
twins methods to control for those unobserved factors that are shared completely by identical twins.
Estimates based on relatively small samples for the US, as well as some larger samples for other countries,
suggest that causal impacts of schooling on health outcomes and behaviors are insignificant or much smaller
than suggested by cross-sectional associations. This study contributes new estimates of cross-sectional
associations and within-MZ causal effects of twins using three relatively large US samples: Mid-Atlantic Twin
Registry, Minnesota Twin Registry and NAS-NRC Twin Registry of WWII Military Veterans. The estimates
suggest that schooling is significantly associated with numerous health outcomes and behaviors in the US.
However if within-MZ twins estimators are used to control for unobserved factors, there is no causal
relationship between schooling and better health behaviors. There is some evidence that more schooling
causally affects self-reported health and overweight status, net of unobservable cofounders, though not to the
extent that cross-sectional associations suggest. Finally, spousal schooling is associated with better health
outcomes and behaviors, but there is no evidence of any causal effect.
Keywords
Schooling, Education, Twins, Adult Health, Cross-Sectional Data, Mid-Atlantic Twin Registry, Minnesota
Twin Registry, NAS-NRC Twin Registry of WWII Military Veterans, Health Outcomes, Schooling
Attainment
Disciplines
Demography, Population, and Ecology | Social and Behavioral Sciences | Sociology
Comments
Amin, Vikesh, Jere R. Behrman, and Hans-Peter Kohler. 2013. "Schooling has smaller or insignificant effects
on adult health in the US than suggested by cross-sectional associations: New estimates using relatively large
samples of identical twins." Philadelphia, PA: Population Studies Center, University of Pennsylvania. PSC
Working Paper Series, PSC 13-12. http://repository.upenn.edu/psc_working_papers/52
This working paper is available at ScholarlyCommons: http://repository.upenn.edu/psc_working_papers/52
Schooling has smaller or insignificant effects on adult health
in the US than suggested by cross-sectional associations:
New estimates using relatively large samples of identical twins
Vikesh Amina
Jere R. Behrmanb,
Hans-Peter Kohlerc
9 December 2013
Abstract: Adult health outcomes and health behaviors generally are strongly associated
with schooling attainment. But such associations do not necessarily imply that schooling
has causal effects on health outcomes and behaviors of the magnitudes of the
associations. Schooling may be proxying for unobserved factors that are related to
genetics and family background. Recently several studies have used within-identical
(monozygotic, MZ) twins methods to control for those unobserved factors that are shared
completely by identical twins. Estimates based on relatively small samples for the US, as
well as some larger samples for other countries, suggest that causal impacts of schooling
on health outcomes and behaviors are insignificant or much smaller than suggested by
cross-sectional associations. This study contributes new estimates of cross-sectional
associations and within-MZ causal effects of twins using three relatively large US
samples: Mid-Atlantic Twin Registry, Minnesota Twin Registry and NAS-NRC Twin
Registry of WWII Military Veterans. The estimates suggest that schooling is
significantly associated with numerous health outcomes and behaviors in the US.
However if within-MZ twins estimators are used to control for unobserved factors, there
is no causal relationship between schooling and better health behaviors. There is some
evidence that more schooling causally affects self-reported health and overweight status,
net of unobservable cofounders, though not to the extent that cross-sectional associations
suggest. Finally, spousal schooling is associated with better health outcomes and
behaviors, but there is no evidence of any causal effect.
a
Corresponding Author. Assistant Professor, Department of Economics, Central Michigan
University, Mount Pleasant, Michigan, MI 48859, USA, [email protected]. Tel:1-989-7743372. Fax: 1-989-774-2040
b
William R. Kenan, Jr. Professor of Economics and Sociology, McNeil 160, 3718 Locust Walk,
University of Pennsylvania, Philadelphia, PA 19104-6297, USA, [email protected].
c
Fredrick J. Warren Professor of Demography, McNeil 272, 3718 Locust Walk, University of
Pennsylvania, Philadelphia, PA 19104-6297, USA, [email protected]
Acknowledgements: The National Institute of Child Health and Development grant RO1
HD046144-01 on “Causal Effects of Schooling on Adult and Child Health” provided background
support for this study.
1
Introduction
It is well-documented that more-schooled individuals tend to have better health outcomes
and health behaviors. Economic theory posits that a causal relationship may exist, because moreschooled individuals are more efficient in health production (Grossman 1972), that is, the
process by which individuals use health inputs like time, money and behaviors to “produce”
health over the life-course. The productive efficiency theory hypothesizes that more-schooled
individuals produce more health from a given set of inputs. The allocative efficiency hypothesis
states that schooling makes individuals aware of the negative effects of certain health behaviors
and more efficient at using health information. More-schooled individuals also have higher
incomes, and are able to afford better health care.
However, associations between schooling and health do not necessarily reflect causal
relationships that imply that increased schooling improve individuals’ health. Instead the
associations could be cofounded by unobserved factors such as ability that jointly affect
schooling and health. For example, Cutler and Lleras-Muney (2010) found that 20 percent of the
schooling health gradient is driven by cognitive ability. Moreover, there is also possibly reverse
causation, insofar as health (particularly early life health) affects schooling.
Recent studies have attempted to estimate causal effects of schooling in the US using
instrumental variable (IV) estimates based on natural policy experiments regarding schooling.
Some studies have instrumented college completion of men using variation in college attainment
induced by draft avoidance behavior during the Vietnam War. These studies have found college
completion reduced mortality, smoking, body mass index (BMI) and the probability of being
overweight (De Walque 2007, Grimard and Parent 2007, MacInnis 2006, Buckles et al. 2012).
Other studies have used changes in compulsory schooling laws as an instrument and found that
more schooling improved self-reported health (Mazumder 2008), old age cognitive functioning
(Glymour et al. 2008), reduced BMI and probabilities of being overweight (Grabner 2008), but
had no effects on hospitalizations, functional limitations and specific health conditions
(Mazumder 2008). Using changes in compulsory schooling laws Lleras-Muney (2005) found that
more schooling reduced mortality, but Mazumder (2008) showed these results were not robust to
inclusion of state-specific trends.
An alternative to the IV strategy using policy changes is to use data on MZ
(monozygotic; identical) twins and within-MZ twin estimators. MZ twins share the same genetic
makeup and family-rearing environment. By relating twin-pair differences in health measures to
twin-pair differences in schooling, schooling effects on health can be estimated controlling for
influences of unobserved factors that affect schooling and health of both twins. Fujiwara and
Kawachi (2009) and Lundborg (2013) applied this strategy using data for 694-701 MZ twins
from the 1995 Midlife in the United States Survey (MIDUS). Both studies found no significant
effects of schooling on health outcomes (self-reported health, perceived mental health, number of
chronic conditions, BMI) and health behaviors (smoking, exercise). Lundborg (2013) did find
significant postive causal impacts of schooling on self-reported health and negative causal
impacts on number of chronic conditions, when using dichotomous schooling indicators for high
school, some college and college degree instead of a continous measure of grades of schooling.
Other studies have been published on other countries. For example, Behrman et al. (2011) used
this approach to investigate relations between schooling and hospitalization and mortality using
data on 5,294 MZ twins in Denmark. They found significant cross-sectional associations, but no
significant effects of schooling with within-MZ twin estimators.
2
This study provides more evidence of the schooling health-gradient using within-MZ twins
methodology and much larger US twin datasets than in MIDUS from the Mid-Atlantic Twin
Registry, Minnesota Twin Registry and the NAS-NRC Twin Registry of WWII Military
Veterans. These new explorations with larger sample sizes permitted assessment of whether the
mostly insignificant results reported in the above studies for the US were in part the artifact of
relatively small sample sizes. We found that more schooling was associated with better health
outcomes and behaviors. With control for unobserved factors, through within-MZ twins
estimators, there was no causal relationship between schooling and better health behaviors. We
found, however, some evidence that more schooling was associated with better health (selfreported health and overweight status), net of unobserved cofounders, though not to the extent
that the cross-sectional associations suggested. Finally, we investigated whether there were
positive spillover effects of spousal schooling on health. We found that spousal schooling was
associated with better health outcomes and behaviors, but no evidence of any causal effects.
Datasets
The Mid-Atlantic Twin Registry (MATR)
The MATR is a population-based registry of twin pairs ascertained from birth records
and school system records of Virginia, North Carolina, and South Carolina (see Anderson et al.
2002 for details). We used data from the Virginia 30K MATR sample. The sample was
ascertained from public birth records in Virginia for twins born in 1915-1972 and responses to a
letter published in the newsletter of the American Association of Retired Persons. Twins were
mailed health and lifestyle questionnaires in 1987. Completed questionnaires were obtained from
14,763 twins, approximately 70 percent. Our analysis was based on samples of white MZ twin
pairs aged 25-75, with non-missing information on self-reported schooling, reports on the cotwin’s schooling, and health measures. The sample sizes ranged from 3,392 to 4,328.
Twins were asked to report their own schooling, their co-twin’s schooling, and their
spouse’s schooling. Schooling attainment was measured using a 6 point scale: 0-7 grades of
elementary school; 8 grades of elementary school; 1-3 grades of high school; 4 grades of high
school; 1-3 years of college; 4 or more years of college. These categories were recoded as 5, 8,
10, 12, 14 and 16 schooling grades, respectively.
Self-reported health was measured by “how would you rate your health in the past 12
months” with responses ranging from 1=very poor, 2=poor, 3=average, 4=good and 5=very
good. BMI was calculated as (weight in kg)/(height in meters)2, where both height and weight
were self-reported. A twin was defined as being overweight (obese) if his/her BMI was greater or
equal to 25 (30). Never smoked was measured through an indicator equal to 1 if the twin had
never smoked and 0 otherwise. Twins were asked how many alcoholic drinks they consumed in a
typical week with options for 0, 1-3, 4-6, 7-12, 13-18, 19-24, 25-42 and more than 42 drinks per
week. We recoded these categories as 0, 2, 5, 8.5, 15.5, 21.5, 33.5 and 53 respectively. Twins
were asked to describe their leisure exercise with options for no exercise; occasional exercise;
regular exercise about once a week; exercise a couple of time a week or jogging, cycling to work
or vigorous sport activity at least 3-4 times a week. Exercise was measured through a
dichotomous variable equal to 1 if the twin reported regular exercise once a week, exercise a
couple of times a week or jogging, cycling to work or vigorous sport activity at least 3-4 times a
week and 0 otherwise.
3
The Minnesota Twin Registry (MTR)
The MTR contains birth records on all twins born in Minnesota in 1936-1955 (details are
in Lykken et al. 2002). We used data from the Socioeconomic Survey of Twins, which was a
questionnaire sent in 1994 to 6638 members of same sex twin pairs. There was a 55 percent
response rate, with 3680 valid questionnaires returned. Our analysis was based on samples of
white MZ twin pairs, with non-missing information on self-reported schooling, co-twin’s
schooling, and health measures. The sample sizes ranged from 1,290 to 1,310.
Twins were asked to report their own schooling, their co-twin’s schooling, and their
spouse’s schooling. The questionnaire contained several questions pertaining to schooling
attainment: (1) highest grade 1 through 12 of schooling completed; (2) vocational schooling (3)
college schooling (4) graduate/professional schooling and (5) schooling at time of first marriage
and current marriage. We first determined the highest qualification obtained and assigned actual
grades of schooling if no high school diploma, 11 if GED, 12 if high school diploma, 13 if
vocational diploma, 14 if associate degree, 16 if college degree, 18 if masters degree, 19 if JD or
MBA and 20 if doctoral degree.
The questionnaire contained a limited number of health questions. Twins were asked
“how would you rate your health at the present time?” with responses for 1=bad, 2=poor, 3=fair,
4=good, 5=excellent. BMI was calculated as weight in kg/height in meters2, where both height
and weight were self-reported. Twins were defined as being overweight (obese) if their BMI was
greater or equal to 25 (30). No information on health behaviors was available.
NAS-NRC Twin Registry of WWII Military Veterans
The NAS-NRC twin registry consists of 15,924 white male twins born in 1917-1927,
both of whom served in the armed forces. Twins were mailed questionnaires in 1967, 1983 and
2000. We used data from the 2000 survey, which contained information on both health and
schooling. 2060 twin pairs responded to the questionnaire (Page 2002). Our analysis was based
on samples of white MZ twin pairs, with non-missing information on self-reported schooling,
and health measures. The sample sizes ranged from 1726 to 1902.
The questionnaire asked twins their own schooling, but not their co-twins’ or spouses’
schooling. Schooling attainment was measured through a single question asking “what is your
highest grade level, diploma, degree completed?”, with options for none, 1 grade, 2 grades, 3
grades, 4 grades, 5 grades, 6 grades, 7 grades, 8 grades, 9 grades, 10 grades, 11 grades, high
school or GED, 1 year of trade/vocational school after high school, 1 year of college, 2 years of
college or associates degree, 3 years of college, bachelors degree, 1 year of graduate work,
masters degree, some doctoral work, doctoral degree. We assigned actual grades if high school
was not completed, 12 grades for graduating from high school, 13 grades for 1 year of trade
school, 13 grades for 1 year of college, 14 grades for 2 years of college/associate degree, 15
grades for 3 years of college, 16 grades for a bachelors degree, 17 grades for 1 year of graduate
work, 18 grades for a masters degree, 19 grades for some doctoral work and 20 grades for a
doctoral degree.
Self-reported health was measured by “in general would you say your health is” 1=poor,
2=fair, 3=good, 4=very good, 5=excellent. BMI was calculated as weight in kg/height in meters2,
where both height and weight were self-reported. A twin was defined as being overweight
(obese) if his/her BMI was greater or equal to 25 (30). Twins were asked whether their health
limited (1) vigorous activities, (2) moderate activities, (3) lifting or carrying groceries, (4)
climbing several flights of stairs, (5) climbing one flight of stairs, (6) bending, kneeling or
4
stooping, (7) walking more than one mile, (8) walking several blocks, (9) walking one block and
(10) bathing or dressing yourself. The questions had 3 possible responses- 1=no, not limited at
all, 2=yes limited a little, 3=yes, limited a lot. We created an overall measure encapsulating how
much health limited activities by summing responses to the 10 questions. For specific health
conditions, we examined dichotomous variables indicating whether twins have ever had a heart
attack, a stroke, diabetes, high blood pressure, prostate cancer. As a final health outcome, we
examined the number of health problems that twins had in old age. Specifically the twins were
asked whether they had any of the following problems: abdominal or aortic aneurysm, brain
aneurysm, rheumatoid arthritis, diverticulitis, emphysema, hemorrhoids or piles, hiatal hernia,
kidney stone liver cirrhosis, prostate enlargement, cataracts, glaucoma, macular hole in retina,
macular degeneration. For health behaviors, we used an indicator variable for having never
smoked and an indicator variable equal to 1 if the twin reported that they drank beer daily, 3-6
times a week, or twice a week.
Statistical Analysis
To illustrate the within-MZ twins approach, suppose that the health measure of twin i in
pair j (Yij) is related to schooling (Sij), unobserved factors (such as ability, time preference,
childhood family characteristics) that are common to both MZ twins in a given pair (µj) and an
unobserved stochastic term (εij).
Yij = βSij + µj + εij
(1)
An OLS regression of relation (1) provides an estimate of the association between
schooling and health, which is a biased estimate of the causal impact for two reasons. First, β is
likely to be biased because schooling is partially related to unobserved factors that affect both
schooling and health directly (with the bias upward if the unobserved factors affect both
schooling and health in the same direction). Second, classical measurement error in schooling
causes β to be downward-biased. To control for measurement error, self-reported schooling can
be instrumented with the co-twin’s report (or some other report, as by the twins’ children in
Behrman et al. 1994) of the other twin’s schooling. The unobserved factors influences can be
controlled through the within-MZ twin estimator.
Y1j – Y2j= β(S1j – S2j ) + (ε1j - ε2j)
(2)
The influence of unobserved factors µj is differenced out as MZ twins are genetically identical,
the vast majority of MZ twins grow up together, and share many other socioeconomic contexts
such as parental families, schools and neighborhoods.
There are two problems with within-MZ twins estimators in relation (2) that may still
lead to biased estimates. First, the estimates are based on MZ twin pairs that differ in schooling
attainment. If differences in schooling attainment are due to factors that also directly affect
health, then the estimates will be biased (Bound and Solon 1999, Behrman et al. 2011, Kohler et
al. 2011). Specifically β is upward- (downward-) biased if the factor leading to differences in
schooling attainment affects schooling and health in the same (opposite) direction (see Behrman
et al. 2011, Kohler et al. 2011).
For example, schooling differences between MZ twins may be due to differences in youth
health, that also directly affect later life health. Both the MATR and MTR surveys asked twins
whether they have ever suffered from any health conditions such as heart failure, high blood
5
pressure, asthma, hearing impairment, depression, and if so at what age these conditions were
diganosed. We attempted to control for early life health differences as a source of bias, by
including an indicator equal to one if the twin was diagnosed with a health condition at age 16 or
younger in all cross-sectional and within-MZ twins regressions. Second, the attenuation bias due
to measurement error is exacerbated by first differencing. To deal with measurement error, we
followed the Ashenfelter and Krueger (1994) strategy of instrumenting the difference in selfreported schooling with the difference in the co-twin’s report of the other’s schooling.
Relation (4) was used to estimate the effect of spousal schooling on own health. Relation
(4) is the same as relation (1), but also includes the schooling of the spouse of twin i in pair j
(Spij).
Yij = βSij + δSpij + µj + εij
(4)
The cross-spouse effect δ is a biased estimate of the causal effect, because of the unobserved
factors that are correlated with spousal schooling and also directly affect own health. For
example, high-ability individuals tended to have better-schooled spouses and better own health.
The bias due to unobserved factors is differenced out by contrasting the health of MZ twins with
different spousal schooling in relation (5), and approach used for the first time to our knowledge
in this and the concurrent study by Behrman et al. (2013).
Y1j – Y2j= β(S1j – S2j ) + δ(Sp1j – Sp2j ) + (ε1j - ε2j)
(5)
Because within-MZ schooling differences exist over most schooling levels, MZ within
estimates are likely to be closer to average treatment effects (ATE) than local average treatment
effects (LATE) (Moffitt 2009). IV-approaches that rely on variation in compulsory schooling as
a first-stage instrument to predict schooling (Angrist and Krueger 1991; Lleras-Muney 2005), for
example, yield local average treatment effects (LATE) relevant for individuals who are at the
margin to be affected by the instruments used (e.g., the margin of completing only compulsory
schooling levels) but not average treatment effects for broader populations beyond this margin
(Behrman et al. 2011, Kohler et al. 2011, Lundborg 2013). ATE (or approximations thereof) that
are provided by within-MZ approaches are likely preferable to LATE estimates that result from
the use of such instruments as compulsory schooling regulations.
Results
Table 1 provides descriptive statistics. The average age of twins at the time of survey was
46 in the MTR, 52 in the MATR and 74 years in the NAS-NRC. The majority of twins in the
MTR and MATR were female. The average grades of schooling was similar in the MTR and
NAS-NRC datasets (approximately 14 grades), and lower in the MATR (13.52 grades). Only 2
percent of twins did not graduate from high school in the MTR, compared to 9 percent and 12
percent in the MATR and NAS-NRC. The proportion of twins that graduated from high school,
had some post-high school schooling, and had bachelors degrees was similar across datasets.
Although the MATR had the largest sample of twins, it had the least variation in grades of
schooling. On average twin pairs had a difference of 0.69 grades and 71 percent of twin pairs had
no difference in grades. In comparison in the MTR and NAS-NRC, 51 percent and 43 percent of
twin pairs had the same grades of schooling, and on average the twin pairs had differences of
1.10 and 1.40 grades.
6
Average self-reported health was similar in the MATR and MTR (4.34 and 4.37) but
substantially lower in the NAS-NRC (3.48). Average BMI was similar in all datasets.
Approximately one half of twins were overweight in the MATR and NAS-NRC, whereas in the
MTR only 38 percent of twins were overweight. In the MTR 16 percent of twins had an early
health problem, more than double the proportion in the MATR.
There was also fairly substantial within-twin pair variation in the health measures. The
absolute within-twin pair difference in self-reported health was 0.48 in the MTR, 0.58 in the
MATR and and much higher in the NAS-NRC (0.84). In all three datasets, the absolute withintwin pair difference in BMI was over 2 units, and over 18 (8) percent of twin pairs were
discordant on overweight (obesity) status. 20 percent of twin pairs differed on smoking status in
the MATR and NAS-NRC. The absolute within-twin pair diffference in number of health
problems for the elderly NAS-NRC twins was 1.22. For specific health conditions such as ever
had a diabetes, prostrate cancer, heart attack, the proportion of twin pairs that differed was 11, 15
and 22 percent respectively.
The main results are in Table 2. All cross-sectional regressions control for age, age
squared as well as gender and early health problems in the MATR and MTR. The estimates in
column 1 indicate that schooling was associated with better health outcomes in the three data
sets, in terms of higher self-reported health, lower BMI and lower probability of being
overweight and obese. Moreover, in the elderly sample of NAS-NRC twins, more schooling was
associated with having fewer health-limiting problems. However, there was no relationship
between schooling and specific health conditions (heart attack, diabetes, high blood pressure,
prostate cancer) and the total number of health problems in old age. Schooling was associated
with better health behaviors. More-schooled individuals had higher probabilities of having never
smoked, and of undertaking some exercise in leisure time. More schooling was associated with a
lower probability of drinking beer frequently in the NAS-NRC. Surprisingly, however, more
schooling was associated with more alcohol consumption in the MATR. The positive association
may be plausible if people believed that there were health benefits of modest alcohol
consumption or if alcohol was more affordable to more-schooled individuals due to higher
income. These associations increased in magnitude, once measurement error was controlled
(column 2) by instrumenting self-reported schooling with reports from co-twins.
Column 3 shows within-MZ-twins estimates, which also controlled for the early health
problems indicator in the MATR and MTR. Compared to the cross-sectional estimates in column
1, the majority of the within-MZ twins estimates were substantially smaller in magnitude and
statistically insignificant. This suggests that there was no causal effect of schooling. Rather the
cross-sectional associations were cofounded by unobserved factors that affected schooling and
health directly. There were however two exceptions. First, schooling was still significantly
associated with better self-reported health and lower probabilities of being overweight in the
MATR, even when controlling for unobserved factors in column 3. An extra grade of schooling
increased self-reported health by 0.05 units, on a scale of 0-5, and lowered probabilities of being
overweight by 1.2 percent. When self-reported schooling was instrumented using co-twins’
reports of the others’ schooling to control for random measurement error in column 4, the point
estimates increased to 0.097 and 4.7 percent respectively, although the former estimate was no
longer significant. Second, schoolings appeared to positively impact self-reported health in old
age, though only about half as much as the cross-sectional association in column 1. An extra
grade of schooling increased self-reported health by 0.035 units, on a scale of 0-5 for elderly men
in the NAS-NRC sample. The point estimates for the other health measures also increased in
7
absolute magnitudes, when instrumenting for measurement error in column 4. They all remained
statistically insignificant, with the exception of exercise in the MATR, which was significant at
the 10 percent level.
Appendix table 1 reports results by gender for MTR and MATR, and table 2 summarizes
the difference in schooling coefficients for women and men. The cross-sectional analyses again
indicated that more schooling was associated with better health outcomes and behaviors, with the
possible exception of alcohol consumption. For self-reported health, the coefficient on the “early
health problem indicator” was negative and significant, indicating that twins that suffered from a
health condition in youth reported worse health compared to those who did not suffer from any
health problems. There were only two significant gender differences in the cross-sectional
schooling coefficient estimates: larger absolute magnitudes of associations of an extra grade of
schooling with BMI and with never smoked for women than for men in the MATR sample. But
neither of these gender differences was significant in the within-MZ estimates. On the other
hand, every additional grade of schooling reduced probabilities of being overweight in the MTR
sample by 5.6% more for men than for women. In the MATR, to investigate heterogeneous
effects by age we interacted grades of schooling with indicator variables for being (1) 25-44, (2)
45-64 and (3) 65-75 years of age. There was no significant differential effect of schooling by
these age groups (see appendix table 3).
As noted above, IV estimates reflect LATEs. For example, IV estimates based on
changes in compulsory schooling laws estimate effects of additional grades of schooling at the
low ends of schooling distributions. In contrast, because within-MZ twin schooling differences
are across schooling distibutions, we estimated effects of both low and high levels of schooling
with non-linear specifications for schooling as in table 3. (No measurement-error-corrected IV
estimates are presented because with binary variables, measurement error is non-classical:
individuals in the lowest educational category cannot under-report their education and
individuals in the top categories cannot over-report their education.) Non-linearities were
introduced through dichotomous variables indicating: (1) whether graduated from high school or
lower, (2) whether had some post-high school schooling and (3) whether had bachelors degree
or higher. The cross-sectional results indicate that those with some post-high school schooling
and those with bacelors degrees had better health outcomes and behaviors compared to those
who have graduated from high school or lower. The coefficient on having bachelors degrees or
higher was almost twice the coefficient on some post-high school schooling, suggesting that
there were larger health returns from completing college. Controlling for unobserved factors
using within-MZ twins estimators (column 3) produced estimates that generally were smaller in
magnitude and statistically insignificant. As before there was still some indication that schooling
led to better self-reported health and lower probabilities of being overweight in the MATR.
Individuals with bachelors degrees or higher had higher (lower) self-reported health
(probabilities of being overweight) by 0.257 units (6.4 percent) compared to individuals who had
graduated high school or lower. In the NAS-NRC sample, there was some evidence that the
representation of schooling matters for the obtained results. Using linear grades of schooling
(table 2), an extra grade of schooling led to higher reported health. However, using a non-linear
representation, the within-MZ twin estimates for self-reported health were insgnificant.
Moreover, the within-MZ twins estimates for BMI, suggested that men who had at least
bachelors degrees had lower BMI of 0.574 units compared to men who graduated high school or
lower. In comparison, there was no significant association of grades of schooling and BMI
within-MZ twin pairs in table 2.
8
Table 4 reports estimates of spousal schooling on own health. All cross-sectional
estimates in columns 1 and 2 controlled for age, age squared, gender and early health problems.
The estimates in column 1 show that spousal schooling was associated with better self-reported
health, lower BMI and lower probabilities of being overweight and obese. Spousal schooling
also was associated with increased probabilities of having never smoked, undertaking some
exercise in leisure time and alcohol consumption. The absolute magnitudes of estimated
associations with spousal schooling fall in column 2, when own schooling was introduced.
Spousal schooling still had positive associations with health outcomes, but no longer had any
significant associations with health behaviors. For health outcomes in the MTR, it was surprising
that the associations with spousal schooling were larger than those with own schooling. One
would usually expect the opposite, which was found in the MATR. The within-MZ twins
estimates in columns 3 and 4 were all insignificant, suggesting that the associations in columns 1
and 2 were due to influences of unobserved factors. Appendix table 4 shows estimates of
interactions between the female indicator and (1) spousal and (2) own schooling in crosssectional and within-MZ twins regressions. The cross-sectional regressions indicated that
husband’s schooling had greater effects in reducing BMI, probabilities of being overweight,
obese in the MTR, and increasing probabilities of never smoking, engaging in frequent exercise
and alcohol consumption compared to wives’ schooling in the MATR. None of these differences
were statistically significant when controlling for unobserved factors within-MZ twin pairs. In
the MATR there were significant differential effects of spousal schooling on BMI and
probabilities of being overweight, within-MZ twin pairs. An extra grade of wife’s schooling
reduced BMI by 0.155 units and the probability of being overweight by 1.8 percent compared to
husband’s schooling. In the MTR, wife’s schooling increased self-reported health by 0.53 units
compared to husband’s schooling, within-MZ twin pairs.
Discussion
This study uses large US twins datasets and found that more schooling generally was
associated with better health outcomes and health behaviors with the possible exception of more
alcohol consumption. The within-MZ twins estimates indicated that more schooling did not
causally improve health behaviors. Rather the positive cross-sectional relationships appeared to
be reflecting influences of unobserved factors that affected schooling and health behaviors
directly. This finding is in line with previous US twin-based evidence (Fujiwara and Kawachi
2009, Lundborg 2013), as well as studies of other countries (Behrman et al. 2011). We found,
however, some evidence that more schooling causally led to some better health outcomes,
though generally with smaller effects than suggested by cross-sectional associations. More
schooling was still significantly associated with better self-reported health and lower
probabilities of being overweight, once unobserved factors were controlled with within-MZ twin
estimators. The beneficial effect was mainly driven by completing college. Like Lundborg
(2013), we also found the functional form of schooling can lead to different results. Using linear
grades of schooling, the within-MZ twins estimates showed no association between schooling
and BMI for old-age men. In contrast, using a non-linear functional form, we found that old-age
men with bachelor degrees had lower BMI compared to old-age men who had graduated from
high school or lower. Finally, spousal schooling was associated with better health outcomes and
behaviors, but the associations disappeared once unobserved cofounders were removed withinMZ twin pairs. We also did not find any consistent evidence for gender differences in effects of
own schooling and spousal schooling.
9
This study has several limitations. First, within-MZ twins estimates may still be biased if
differences in schooling attainment were due to factors that also directly affected health. We
attempted to control for one possible factor- early life health differences. Differences in
schooling attainment may be due to other factors such as differences in parental treatment or
school quality, that also have direct affects on health, which we cannot rule out. But if these
factors were positively associated with both schooling and health, the fact that we cannot control
for them led our within-MZ estimates to be upper bounds on absolute magnitudes of true
schooling effects. Because our results in many cases indicated no or only weak effects of
schooling on health, our basic conclusion that the causal effects of schooling on health are
generally weak or non-existent is not affected by such biases. Second, we used data on twins
born in specific states and/or from certain birth cohorts, who were therefore unlikely to have
been representative of the general US population. To assess the representativeness of the twins
data, we compared the MATR (MTR) twins to nationally-representative samples of white
individuals aged 25-75 (39-58) in the 1987, (1994) Behavioral Risk Factor Surveillance System
(BRFSS). The elderly male NAS-NRC twins were compared to elderly male veterans in the 2000
BFRSS. Summary statistics from the BFRSS in appendix table 5 show that the twin samples
tended to have higher schooling attainment than the nationally-representative samples from these
birth cohorts. The MATR and MTR twins data sets also over-represented women. Average selfreported health was lower in the 1994 and 2000 BFRSS compared to the MTR and NAS-NRC
twins. BMI was similar in the BFRSS and twins data sets. The proportion of individuals who had
never smoked was slightly lower in the 1987 and 2000 BFRSS compared to the MATR and
NAS-NRC twins. Appendix table 6 presents cross-sectional schooling estimates, where we
combined the twins and BFRSS data and included an interaction between grades of schooling
and a dummy variable for the twins data to test whether schooling coefficients differed in the
twins and BFRSS data sets. In the combined MATR-BFRSS data set, there was a significant
difference in schooling coefficients on BMI, probabilities of being overweight, obese for twins
and individuals in the BFRSS. There was no significant difference for probabilities of having
never smoked. In the combined MTR-BFRSS data set, for BMI and the incidence of overweight
and obesity, the interaction was insignificant, but significant for self-reported health. In the
combined NAS-NRC BFRSS dataset, the interaction was significant only for probabilities of
never smoking. In general, although the results may not be fully generalizable, differences
between the twins and BFRSS samples did not cause biases in our within-MZ twins estimates if
these differences from representative samples were due to the unobserved factors that were
controlled for in the within-MZ estimates. Third, we did not investigate possible mechanisms
under which schooling improved health except that if we included income in our estimates the
basic results for schooling were not altered much (estimates available from authors), suggesting
that schooling was not working primarily through income.
Despite these limitations, this study provides more evidence on causal effect of schooling
on health, to a literature where studies for the US and other developed countries either found no
or a small causal effect of schooling. Our first finding that schooling improved somewhat selfreported health, though not as much as indicated by cross-sectional associations, is consistent
with Lundborg (2013) but contrasts with instrumental variable estimates for the US (Mazumder
2008), UK (Clark and Royer 2013), and Denmark (Arendt 2005), where no effect was found.
Our second finding that more schooling lowered probabilities of being overweight, differed from
within-MZ twins estimates for the US where no causal relationship was found (Fujiwara and
Kawachi 2009, Lundborg 2013), and for Australia where schooling reduced probabilities of
being overweight for men but not for women (Webbink et al. 2010). Instrumental variables
10
estimates for the UK (Clark and Royer 2013), Denmark (Arendt 2005) also showed no evidence
of any causal relationship. However, our finding is consistent with instrumental variable
estimates for the US (Grabner 2008, MacInnis 2006) and Germany (Kemptner et al. 2011).
References
Anderson, L. S., Beverly, W. T., Corey, L. A., & Murrelle, L. (2002). The Mid-Atlantic
Twin Registry. Twin Research, 5(5), 449–455.
Angrist, J. D. & Krueger, A. B. (1991). Does Compulsory School Attendance Affect
Schooling and Earnings. Quarterly Journal of Economics, 106(4), 979-1014.
Arendt, J. N. (2005). Does education cause better health? A panel data analysis using
school reforms for identification. Economics of Education Review, 24(2), 149-160.
Ashenfelter, O., & Krueger, A. (1994). Estimates of the economic return to schooling
from a new sample of twins. American Economic Review, 84(5), 1157–1173.
Behrman, J.R., Kohler, H-P., Jensen, V.M., Pedersen, D., Petersen, I., Bingley, P., &
Christensen, K. (2011). Does schooling reduce hospitalization and delay mortality? New
evidence based on Danish twins. Demography. 48(4), 1347-1375.
Behrman, J.R., Rosenzweig, M.R. & Taubman, P. (1994). Endowments and the
Allocation of Schooling in the Family and in the Marriage Market: The Twins Experiment.
Journal of Political Economy, 102(6), 1131-74.
Behrman, J.R., Xiong, Y., & Zhang, J. (2013). Schooling-Health Associations
Misrepresent the Causal Effects of Schooling on Health: Evidence from Chinese Twins. Hong
Kong: Chinese University of Hong Kong.
Bound, J., & Solon, G. (1999). Double trouble: on the value of twins-based estimation of
the returns to schooling. Economics of Education Review, 18(2), 169-182.
Buckles, K., Ofer, M., Morril, M.S. & Wozniak, A. (2012). The effect of college
education on health. IZA Discussion Paper Number 6659.
Clark, D., & Royer, H. (2013). The Effect of Education on Adult Mortality and Health:
Evidence from Britain. American Economic Review, 103(6), 2087-2120.
Cutler, D. M. & Lleras-Muney, A. (2010). Understanding differences in health by
education. Journal of Health Economics, 29, 1-28.
De Walque, D. (2007). Does education affect smoking behaviors?: Evidence using the
Vietnam draft as an instrument for college education. Journal of Health Economics, 26(5), 877895.
Grimard, F., & Parent, D.(2007). Education and Smoking: Were Vietnam War Draft
Avoiders Also More Likely to Avoid Smoking? Journal of Health Economics, 26(5), 896-926.
Grossman, M. (1972). On the concept of health capital and demand for health. Journal of
Political Economy, 80(2), 223-255.
Grabner, M. (2008). The Causal Effect of Education on Obesity. Evidence from
Compulsory Schooling Laws. Mimeo, University of California at Davis.
Glymour, M. M, Kawachi, I., Jencks, C. S. & Berkman, L. F. (2008). Does childhood
schooling affect old age memory or mental status? Using state schooling laws as natural
experiments. Journal of Epidemiology Community Health, 62(6), 532–537.
Fujiwara, T., & Kawachi, I. (2009). Is education causally related to better health? A twin
fixed-effect study in the USA. International Journal of Epidemiology, 38(13), 1310-1322.
Lleras-Muney, A. (2005). The Relationship between Education and Adult Mortality in
the United States. Review of Economic Studies, 72(1), 189-221.
11
Lundborg, P. (2013) “The health returns to education- What can we learn from twins?”
Journal of Population Economics, 26(2), 673-701.
Lykken, D., Bouchard, T., McGue, M. & A., T. (1990). The minnesota twin family
registry – some initial findings. Acta Geneticae Medicae et Gemellologiae, 39(1), 35–70.
Kemptner D., Jurges, H. & Reinhold, S. (2011). Changes in compulsory schooling and
the causal effect of education on health: Evidence from Germany. Journal of Health Economics,
30(2), 340-354
Kohler, H.-P.; Behrman, J. & Schnittker, J. (2011). Social Science Methods for Twins
Data: Integrating Causality, Endowments and Heritability. Journal of Biodemography and Social
Biology, 57(1), 88-141.
MacInnis, B. (2006). The Long-Term Effects of College Education on Morbidities: New
Evidence from the Pre-Lottery Vietnam Draft. Mimeo, University of Michigan.
Mazumder, B. 2008. Does education improve health? A reexamination of the evidence
from compulsory schooling laws. Economic Perspectives, 33(2), 2–16.
Moffitt, R. A. (2009). Issues in the Estimation of Causal Effects in Population Research,
With an Application to the Effects of Teenage Childbearing. In Engelhardt, H., Kohler, H-P &
Fürnkranz-Prskawetz, A. (eds.) Causal Analysis in Population Studies, Berlin: Springer Verlag ,
9-29
Page, W.F. (2002). The NAS-NRC Twin Registry of WWII Military Veteran Twins.
Twin Research, 5(5), 493-496
Webbink, D., Martin, N. & Visscher, P. (2010). Does Education reduce the probability of
being overweight? Journal of Health Economics, 29(1), 29-38.
Table 1: Descriptive Statistics
MATR
MTR
12
NAS-NRC
(1)
(2)
(3)
52.41 (15.04)
46.63 (5.49)
74.16 (2.76)
Background
Characteristics
Age
25-34
0.22 (0.40)
35-44
0.12 (0.33)
0.42 (0.49)
45-54
0.12 (0.33)
0.47 (0.50)
55-64
0.29 (0.45)
0.11 (0.31)
65-75
0.26 (0.44)
0.66 (0.47)
76-85
0.30 (0.46)
Female
0.71 (0.45)
0.65 (0.47)
0.00 (0.00)
Grades of schooling
13.52 (2.10)
14.03
14.02 (3.00)
Absolute within-twin
0.69 (1.17)
1.10 (1.52)
1.40 (0.87)
0.71 (0.46)
0.51 (0.50)
0.43 (0.87)
0.19 (0.39)
0.23 (0.42)
difference in grades
of schooling
0 grade difference
1 grade difference
2 grades difference
0.25 (0.43)
0.14 (0.35)
0.14 (0.35)
3 grades difference
0.00 (0.05)
0.05 (0.22)
0.06 (0.24)
4 grades difference
0.04 (0.18)
0.07 (0.25)
0.07 (0.26)
5 grades difference or more
0.01 (0.07)
0.04 (0.18)
0.06 (0.24)
Less than 12 grades
0.09 (0.28)
0.02 (0.12)
0.12 (0.32)
12 grades
0.34 (0.48)
0.33 (0.47)
0.25 (0.43)
13-15 grades
0.26 (0.44)
0.30 (0.46)
0.25 (0.43)
16 grades or more
0.31 (0.46)
0.35 (0.48)
0.38 (0.48)
13.43 (2.12)
13.91 (2.37)
13.57 (2.29)
13.81 (2.47)
0.07 (0.25)
0.16 (0.36)
Co-twin's grades of
schooling
Spouse's grades of
schooling
Early health problem
Health Outcomes
Self-reported health
4.34 (0.86)
4.37 (0.67)
3.48 (0.98)
BMI
24.54 (4.27)
25.74 (4.49)
25.79 (3.31)
13
Table 1 Continued
Overweight
0.38 (0.47)
0.50 (0.50)
Obese
0.10 (0.30)
0.16 (0.36)
Limiting activities
0.58 (0.49)
0.09 (0.29)
14.54 (4.90)
Ever had heart attack
0.16 (0.37)
Ever had stroke
0.09 (0.29)
Ever had high blood pressure
0.46 (0.50)
Ever had diabetes
0.15 (0.36)
Ever had prostrate cancer
0.12 (0.33)
Number of health problems
1.97 (1.48)
Health Behaviors
Never Smoked
0.53 (0.50)
Exercise
0.52 (0.50)
Alcohol Consumption
2.48 (5.81)
0.37 (0.48)
0.36 (0.47)
Absolute within-twin pair
difference in
Self-reported health
0.58 (0.76)
0.48 (0.63)
0.82 (0.78)
BMI
2.08 (2.28)
2.49 (2.44)
2.28 (2.01)
Number of health problems
1.22 (1.05)
Limiting activities
3.84 (4.29)
Alcohol consumption
2.45 (5.68)
Proportion of twin pairs that
differ in
Overweight
0.19 (0.40)
0.23 (0.42)
0.30 (0.45)
Obese
0.08 (0.28)
0.14 (0.34)
0.11 (0.31)
Ever had heart attack
0.22 (0.41)
Ever had stroke
0.13 (0.33)
Ever had high blood pressure
0.26 (0.44)
Ever had diabetes
0.11 (0.31)
Ever had prostate cancer
0.15 (0.35)
Never Smoked
0.20 (0.40)
Exercise
0.33 (0.47)
0.20 (0.40)
0.33 (0.47)
Alcohol consumption
Notes: Standard deviations in parentheses.
14
Table 2: Cross-Sectional and Within-MZ Twins Estimates
Cross-Section
Within-MZ Twins
OLS
IV
OLS
IV
(1)
(2)
(3)
(4)
0.075
0.081
0.05
0.097
(.007)***
(.009)***
(.017)**
(0.064)
[4238]
[4238]
[2164]
[2164]
Self-reported health
MATR
MTR
NAS-NRC
0.045
0.059
0.014
0.018
(0.008)***
(0.012)***
(0.014)
(0.022)
[1310]
[1310]
[655]
[655]
0.067
0.035
(0.008)***
(0.018)*
[1868]
[934]
BMI
MATR
MTR
NAS-NRC
-0.243
-0.286
-0.063
-0.145
(.037)***
(.045)***
(0.060)
(0.215)
[4272]
[4272]
[2136]
[2136]
-0.182
-0.248
0.061
0.035
(0.061)***
(0.091)**
(0.075)
(0.113)
[1290]
[1290]
[645]
[645]
-0.115
-0.056
(0.030)***
(0.047)
[1844]
[922]
Overweight
MATR
MTR
NAS-NRC
-0.027
-0.03
-0.012
-0.047
(.004)***
(.005)***
(.006)*
(.026)*
[4272]
[4272]
[2136]
[2136]
-0.014
-0.023
0.011
0.011
(0.007)**
(0.010)**
(0.010)
(0.019)
[1290]
[1290]
[645]
[645]
-0.014
-0.005
(.004)***
(0.008)
[1844]
[922]
Obese
MATR
MTR
-0.013
-0.016
-0.002
-0.004
(.002)***
(.003)***
-0.005
-0.017
[4272]
[4272]
[2136]
[2136]
-0.009
-0.012
-0.001
-0.002
(0.005)*
(0.007)*
(0.007)
(0.011)
[1290]
[1290]
[645]
[645]
15
Table 2 Continued
NAS-NRC
-0.007
-0.002
(.002)**
(0.004)
[1844]
[922]
-0.003
0.001
(0.003)
(0.006)
[1882]
[922]
-0.005
0.007
(0.003)
(0.005)
[1876]
[938]
-0.007
-0.004
(0.004)
(0.007)
[1882]
[941]
-0.006
-0.004
(0.015)
(0.007)
[1342]
[671]
-0.262
-0.029
(0.042)***
(0.083)
[1726]
[863]
-0.007
-0.031
(0.015)
(0.026)
[1342]
[671]
Ever had heart attack
NAS-NRC
Ever had diabetes
NAS-NRC
Ever had high blood
pressure
NAS-NRC
Ever had prostrate cancer
NAS-NRC
Limiting activities
NAS-NRC
Number of health problems
NAS-NRC
Never Smoked
MATR
NAS-NRC
0.018
0.022
0.022
0.013
(.005)***
(.005)***
(0.007)
(0.027)
[4126]
[4126]
[2063]
[2063]
0.031
0.007
(0.004)***
(0.007)
[1902]
[951]
Exercise
MATR
0.037
0.041
0.01
0.076
(.004)***
(.005)***
(0.010)
(.036)*
[4240]
[4240]
[2120]
[2120]
16
Table 2 Continued
Alcohol Consumption
MATR
NAS-NRC
0.091
0.116
-0.058
-0.347
(.054)*
(0.064)*
(0.120)
(0.534)
[3392]
[3392]
[1696]
[1696]
(0.020)
0.001
(0.005)***
-0.01
[1048]
[524]
Notes: Cross sectional regressions in the MATR and MTR control for age, age squared, gender and early health
problems. Cross-sectional regressions for NAS-NRC control for age and age squared. For cross-sectional IV
estimates, twin 1’s schooling is instrumented by twin 2’s report of twin 1’s schooling and vice versa. All Within-MZ
twins estimates for MATR and MTR control for early health problems. For within-MZ twins IV estimates, the difference
in self-reported schooling is instrumented by the difference in the co-twin’s report of the other’s schooling. Standard
errors clustered by twin pairs in (.). ***significant at 1% ** significant at 5%; * significant at 10%. Sample size in [.]
17
Table 3: Cross-Sectional and Within-MZ Twins Estimates, Non-Linear Schooling
Cross-Section
Within-MZ Twins
OLS
OLS
0.174
0.101
(0.036)***
(0.063)
0.305
0.257
(0.033)***
(0.081)***
[4328]
[2164]
0.124
-0.021
(0.047)***
(0.065)
Self-reported health
MATR
Some college
Degree
MTR
Some college
Degree
0.262
0.061
(0.047)***
-0.078
[1316]
[655]
0.225
0.055
(0.060)***
(0.093)
0.417
0.158
(0.057)***
(0.123)
[1868]
[934]
0.373
-0.232
(0.193)**
(0.203)
-1.06
-0.251
(0.179)***
(0.320)
[4272]
[2136]
-0.356
0.132
(0.375)
(0.337)
-0.918
0.415
NAS-NRC
Some college
Degree
BMI
MATR
Some college
Degree
MTR
Some college
Degree
(0.358)**
(0.406)
[1290]
[645]
-0.557
-0.373
(0.215)**
(0.243)
-0.727
-0.574
(0.210)***
(0.327)*
[1844]
[922]
NAS-NRC
Some college
Degree
18
Table 3 Continued
Overweight
MATR
Some College
Degree
-0.03
-0.032
(0.021)
(0.024)
-0.123
-0.064
(0.020)**
(0.032)**
[4272]
[2136]
-0.031
0.049
(0.037)
(0.041)
-0.062
0.114
(0.037)*
(0.050)**
[1290]
[645]
-0.048
-0.03
(0.031)
(0.044)
-0.069
-0.043
MTR
Some college
Degree
NAS-NRC
Some college
Degree
(0.029)**
-0.055
[1844]
[922]
-0.03
-0.016
(0.013)**
(0.020)
-0.057
-0.021
(0.011)***
-0.025
[4272]
[2136]
-0.028
-0.059
(0.045)
(0.059)
Obese
MATR
Some College
Degree
MTR
Some college
Degree
-0.039
-0.015
(0.023)*
-0.041
[1290]
[645]
-0.055
-0.043
(0.018)***
(0.024)*
-0.053
-0.045
(0.018)***
(0.031)
[1844]
[922]
NAS-NRC
Some college
Degree
19
Table 3 Continued
Limiting Activities
NAS-NRC
Some college
Degree
-0.69
-0.1
(0.315)***
(0.469)
-1.58
0.007
(0.294)***
(0.607)
[1726]
[863]
0.015
0.019
(0.021)
(0.027)
0.097
0.011
(0.022)***
-0.033
[4126]
[2063]
0.041
-0.021
(0.029)
(0.034)
Never Smoked
MATR
Some college
Degree
NAS-NRC
Some college
Degree
0.219
0.053
(0.029)***
(0.045)
[1902]
[951]
0.087
0.002
(0.021)***
(0.037)
0.178
0.067
Exercise
MATR
Some college
Degree
(0.020)***
(0.046)
[4240]
[2120]
0.391
-0.656
(0.282)
(0.438)
0.303
-0.142
(0.277)
(0.617)
[3392]
[1696]
-0.059
0.045
(0.043)
(0.062)
Alcohol consumption
MATR
Some college
Degree
NAS-NRC
Some college
Degree
-0.147
-0.063
(0.038)***
(0.084)
[1048]
[524]
Notes: The omitted schooling category is high school graduate or lower. Cross sectional regressions in the MATR
and MTR control for age, age squared, gender and early health problems. Cross-sectional regressions for NAS-NRC
control for age and age squared. Within-MZ twins estimates for MATR and MTR control for early health problems.
20
Standard errors clustered by twin pairs in (.). ***significant at 1% ** significant at 5%; * significant at 10%. Sample
size in [.]
21
Table 4: Cross-Sectional and Within-MZ Twins Estimates of Spousal Schooling
Cross-Section
Within-MZ Twins
OLS
OLS
OLS
OLS
(1)
(2)
(3)
(4)
0.046
0.022
0.014
0.011
(0.010)***
(0.011)**
(0.012)
(0.012)
Self-reported health
MATR
Spousal schooling
Own schooling
N
0.049
0.025
(0.012)***
(0.024)
1674
1674
837
837
0.045
0.035
0.009
0.007
(0.009)***
(0.010)***
-0.016
-0.016
MTR
Spousal schooling
Own education
N
0.023
0.019
(0.011)**
(0.019)
868
868
434
434
-0.286
-0.197
-0.024
-0.005
(0.050)***
(0.052)***
(0.045)
(0.044)
BMI
MATR
Spousal schooling
Own schooling
N
-0.182
-0.174
(0.064)***
(0.077)**
1644
1644
822
822
-0.256
-0.207
0.079
-0.075
(0.064)***
(0.067)***
(0.058)
(0.060)
MTR
Spousal schooling
Own schooling
N
-0.108
-0.33
(0.073)
(0.990)
846
846
434
434
-0.03
-0.021
-0.005
-0.004
(0.005)***
(0.006)***
(0.006)
(0.006)
Overweight
MATR
Spousal schooling
Own schooling
N
1644
-0.019
-0.006
(0.007)**
(0.011)
1644
822
822
22
Table 4 Continued
MTR
Spousal schooling
-0.021
-0.018
0.002
0.001
(0.007)***
(0.008)**
(0.008)
(0.008)
Own schooling
N
-0.005
0.008
(0.009)
(0.013)
846
846
434
434
-0.015
-0.011
0.000
0.001
(0.004)***
(0.004)***
(0.005)
(0.005)
Obese
MATR
Spousal schooling
Own schooling
-0.009
-0.009
(0.005)*
N
(0.007)
1644
1644
822
822
-0.012
-0.008
-0.002
-0.001
(0.005)
(0.007)
MTR
Spousal schooling
(0.005)**
Own schooling
(0.007)
-0.009
-0.005
(0.006)
(0.010)
N
846
846
434
434
Spousal schooling
0.012
0.004
-0.003
-0.001
(0.006)***
(0.007)
(0.006)
(0.006)
Own schooling
N
0.016
-0.014
(0.008)**
(0.011)
1648
1648
824
824
0.017
0.003
-0.006
0
(0.006)***
(0.006)
(0.009)
(0.009)
Exercise
Spousal schooling
Own schooling
0.029
-0.001
(0.007)***
(0.017)
N
Alcohol
Consumption
1628
1628
814
814
Spousal schooling
0.134
0.083
0.128
0.157
(0.067)
(0.116)
(0.056)**
Own schooling
(0.124)
0.109
-0.256
(0.084)
(0.171)
N
1338
1338
669
669
Notes: Cross sectional regressions control for age, age squared, gender and early health problems. Within-MZ
Twins regressions control for early health problems. Standard errors clustered by twin pairs in (.). ***significant at 1%
** significant at 5%; * significant at 10%.
23
Appendix Table 1: Cross-Sectional and Within-MZ Twins Estimates, By Gender
Men
Women
Cross-Section
Within-MZ Twins
Cross-Section
Within-MZ Twins
OLS
IV
OLS
IV
OLS
IV
OLS
IV
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
0.041
0.041
0.033
0.031
0.049***
0.075
-0.004
0.007
Self-reported
health
MTR: Schooling
(0.012)***
(0.018)**
(0.020)
(0.029)
(0.011)***
(0.017)***
(0.020)
(0.032)
Early health
Problem
-0.111
-0.111
0.002
0.004
-0.152
-0.161
-0.12
-0.13
(0.082)
(0.082)
(0.113)
(0.110)
(0.071)***
(0.070)***
(0.094)
(0.094)
N
464
464
232
232
846
846
423
423
MATR: Schooling
0.076
0.081
0.021
0.128
0.076
0.084
0.062
0.05
(0.011)***
(0.015)***
(0.036)
(0.076)*
(0.009)***
(0.011)***
(0.021)***
(0.097)
-0.248
-0.252
-0.265
-0.265
-0.296
-0.299
-0.092
-0.089
(.118)***
(.111)**
(0.151)*
(0.147)
(0.067)***
(0.067)***
(0.089)
(0.089)
1262
1262
631
631
3066
3066
1533
1533
-0.165
-0.225
-0.04
0.031
-0.201
-0.272
0.148
0.038
Early health
Problem
N
BMI
MTR: Schooling
(0.075)**
(0.114)*
(0.108)
(0.134)
(0.092)**
(0.135)**
(0.103)
(0.183)
Early health
Problem
-0.848
-0.865
-0.048
-0.04
-0.454
-0.428
-0.34
-0.323
(0.421)**
(0.421)**
(0.405)
(0.401)
(0.438)
(0.442)
(0.333)
(0.335)
N
462
462
231
231
828
828
414
414
MATR: Schooling
-0.091
-0.123
0.071
0.187
-0.304
-0.353
-0.121
-0.379
(0.056)*
(0.067)**
(0.143)
(0.342)
(0.048)***
(0.057)***
(0.059)**
(0.280)
0.357
0.376
-0.157
-0.089
0.585
0.608
0.223
0.242
(0.445)
(0.444)
(0.353)
(0.354)
(0.376)
(0.376)
(0.272)
(0.273)
1256
1256
628
628
3016
3016
1508
1508
-0.013
-0.013
-0.019
-0.001
-0.015
-0.031
0.037
0.022
(0.010)
(0.014)
(0.016)
(0.021)
(0.010)*
(0.013)**
(0.011)***
(0.017)
-0.15
-0.145
-0.082
-0.081
-0.038
-0.033
-0.05
-0.047
(0.063)**
(0.063)
(0.068)
(0.069)
(0.048)
(0.048)
(0.050)
(0.046)
N
462
462
231
231
828
828
414
414
MATR: Schooling
-0.022
-0.026
-0.004
-0.01
-0.028
-0.031
-0.015
-0.058
(0.008)***
(0.009)***
-0.014
-0.037
(0.005)***
(0.006)***
(0.007)**
-0.041
0.006
0.009
-0.12
-0.12
0.04
0.041
-0.005
-0.002
(0.063)
(0.063)
(0.072)*
(0.072)
(0.033)
(0.033)
(0.032)
(0.032)
1256
1256
628
628
3016
3016
1508
1508
Early health
Problem
N
Overweight
MTR: Schooling
Early health
Problem
Early health
Problem
N
24
Appendix Table 1 Continued
Obese
-0.009
-0.012
-0.006
-0.001
-0.01
-0.012
0.003
-0.003
(0.008)
(0.011)
(0.010)
(0.015)
(0.006)
(0.009)
(0.009)
(0.015)
-0.065
-0.064
-0.041
-0.04
-0.029
-0.029
-0.033
-0.032
(0.036)*
(0.036)*
(0.039)
(0.039)
(0.035)
(0.035)
(0.036)
(0.036)
N
462
462
231
231
828
828
414
414
MATR: Schooling
-0.006
-0.007
-0.005
-0.012
-0.016
-0.02
-0.001
0.003
(0.004)*
(0.004)
(0.008)
(0.021)
(0.003)***
(0.004)***
-0.006
-0.028
-0.011
-0.011
0.019
-0.19
0.049
0.051
0.047
-0.046
(0.027)
(0.027)
(0.019)
(0.019)
(0.024)**
(0.024)**
(0.026)*
(0.026)*
1256
1256
628
628
3016
3016
1508
1508
0.047
0.057
-0.002
0
0.009
0.01
0.005
0.025
(0.007)**
(0.009)**
(0.013)
(0.031)
(0.006)
(0.007)
(0.008)
(0.043)
0.024
0.016
-0.02
-0.02
-0.074
-0.074
0.046
0.044
(0.061)
(0.061)
(0.054)
(0.060)
(0.036)**
(0.037)**
(0.040)
(0.037)
1188
1188
594
594
2938
2938
1469
1469
0.045***
0.049***
0.029*
0.072
0.035
0.038
0.002
0.095
(0.006)
(0.008)
(0.017)
(0.045)
(0.005)***
(0.006)***
(0.012)
(0.059)
0.031
0.028
0.1
0.092
-0.036
-0.037
-0.013
-0.017
(0.059)
(0.059)
(0.090)
(0.090)
(0.036)
(0.036)
(0.047)
(0.048)
1250
1250
625
625
2990
2990
1495
1495
Alcohol Consumption
0.035
MATR: Schooling
0.029
-0.07
0.045
0.13
0.174
-0.054
-0.748
(0.104)
(0.132)
(0.239)
(0.690)
(0.062)**
(0.072)**
(0.138)
(0.872)
-1.26
-1.26
0.208
0.198
0.296
0.275
0.281
0.318
(0.743)*
(0.741)*
(0.651)
(0.649)
(0.433)
(0.433)
(0.403)
(0.406)
MTR: Schooling
Early health
Problem
Early health
Problem
N
Never Smoked
MATR: Schooling
Early health
Problem
N
Exercise
MATR: Schooling
Early health
Problem
N
Early health
Problem
930
930
465
465
2462
2462
1231
1231
N
Notes: Cross sectional regressions control for age, age squared, and an indicator for early health problems. For
cross-sectional IV estimates, twin 1’s schooling is instrumented by twin 2’s report of twin 1’s schooling and vice versa.
Within-MZ twins regressions control for early health problems. For within-MZ twins IV estimates, the difference in selfreported schooling is instrumented by the difference in the co-twin’s report of the other’s schooling. Standard errors
clustered by twin pairs in (.). ***significant at 1% ** significant at 5%; * significant at 10%. Sample size in [.]
25
Appendix Table 2: Differential Effect of Schooling, Women versus Men
Cross-Sectional
Cross-Sectional
Within-MZ
Twins
Within-MZ
Twins
OLS
IV
OLS
IV
MTR
0.008 (0.016)
0.034 (0.024)
-0.037 (0.028)
-0.024 (0.043)
MATR
0.00 (0.014)
0.003 (0.018)
0.041 (0.037)
-0.078 (0.123)
MTR
-0.036 (0.119)
-0.047 (0.176)
0.188 (0.149)
0.007 (0.226)
MATR
-0.213 (0.074)***
-0.023 (0.088)***
-0.192 (0.155)
-0.566 (0.442)
MTR
-0.002 (0.014)
-0.018 (0.019)
0.056 (0.019)***
0.023 (0.027)
MATR
-0.006 (0.009)
-0.005 (0.010)
-0.011 (0.016)
-0.048 (0.055)
MTR
-0.001 (0.009)
0.00 (0.014)
0.008 (0.014)
-0.002 (0.015)
MATR
-0.010 (0.005)**
-0.013 (0.006)***
0.004 (0.010)
0.015 (0.035)
-0.038 (0.009)***
-0.047 (0.011)***
0.007 (0.015)
0.025 (0.053)
-0.010 (0.008)
-0.011 (0.010)
-0.027 (0.021)
0.023 (0.074)
0.095 (0.121)
0.145 (0.150)
0.016 (0.276)
-0.793 (1.11)
Self-reported
health
BMI
Overweight
Obese
Never smoked
MATR
Exercise
MATR
Alcohol
consumption
MATR
Notes: Standard errors in parentheses. ***significant at 1% ** significant at 5%; * significant at 10%.
26
Appendix Table 3: Heterogeneous Schooling Effects By Age
Cross-Section
Within-MZ Twins
OLS
OLS
0.067
0.053
(0.011)***
(0.029)*
0.023
0.001
(0.016)
(0.043)
-0.002
-0.008
(0.017)
(0.038)
4238
4238
Years of education
-0.224
-0.092
(0.069)***
(0.077)
(age 45-64)*education
-0.006
0.028
(0.093)
(0.146)
(age 65-75)*education
-0.032
0.059
(0.094)
(0.115)
N
4272
2136
-0.022
-0.018
(0.007)***
(0.011)*
-0.006
-0.006
(0.009)
(0.014)
-0.005
0.031
(0.010)
(0.017)*
4272
2136
-0.009
-0.007
(0.004)**
-0.006
(age 45-64)*education
-0.005
-0.012
(0.006)
(0.010)
(age 65-75)*education
-0.006
-0.014
(0.007)
(0.013)
4272
2136
Self-reported health
Years of education
(age 45-64)*education
(age 65-75)*education
N
BMI
Overweight
Years of education
(age 45-64)*education
(age 65-75)*education
N
Obese
Years of education
N
27
Appendix Table 3 Continued
Smoking
Years of education
(age 45-64)*education
(age 65-75)*education
N
0.061
0.002
(0.007)***
(0.014)
-0.052
-0.004
(0.010)***
(0.017)
-0.072
0.007
(0.011)***
(0.019)
4126
2063
0.077
0.114
(0.090)
(0.175)
-0.077
-0.303
(0.134)
(0.292)
0.139
-0.182
(0.120)
(0.257)
3392
1696
0.042
0.004
(0.007)***
-0.017
-0.086
0.009
(0.068)
(0.023)
-0.009
0.012
(0.010)
(0.027)
Alcohol Consumption
Years of education
(age 45-64)*education
(age 65-75)*education
N
Exercise
Years of education
(age 45-64)*education
(age 65-75)*education
N
4240
2120
Notes: Estimates are based on the regression: Yij=bo + b1 Own Schoolingij + b2(Age45to64)ij+ b4(Age65-75)ij
(
b4Femaleij + b6EarlyHealthProblemi + b6 Own Schooling)ij*(Age45to64)ij + b7 ((Own Schoolingi)ij*(Age65to75)ij + uij..
The omitted age group is 24-44. Standard errors clustered by twin pairs in parentheses. ***significant at 1%
**significant at 5% *significant at 10%
28
Appendix Table 4: Differential Effect of Spousal and Own Schooling
Cross-Section
Within-MZ Twins
Spousal schooling
0.03 (0.23)
0.041 (0.028)
Own Schooling
-0.035 (0.025)
0.017 (0.054)
Spousal schooling
-0.007 (0.021)
-0.053 (0.032)*
Own Schooling
0.017 (0.022)
-0.026 (0.039)*
Spousal schooling
-0.167(0.105)
-0.155 (0.085)*
Own Schooling
-0.047 (0.123)
-0.017 (0.157)
Spousal schooling
-0.234 (0.136)*
-0.005 (0.111)
Own Schooling
-0.035 (0.149)
-0.009 (0.201)
Spousal schooling
0.008 (0.015)
-0.008 (0.014)
Own Schooling
-0.005 (0.015)
-0.024 (0.025)
Self-reported health
MATR
MTR
BMI
MATR
MTR
Overweight
MATR
MTR
Spousal schooling
-0.015 (0.017)
0.000 (0.017)
Own Schooling
-0.005 (0.018)
0.053 (0.026)**
Obese
MATR
Spousal schooling
-0.020 (0.008)**
-0.018 (0.011)
Own Schooling
-0.002 (0.009)
0.009 (0.016)
Spousal schooling
-0.024 (0.011)**
-0.016 (0.014)
Own Schooling
0.001 (0.011)
0.007 (0.020)
Spousal schooling
-0.028 (0.014)**
-0.013 (0.014)
Own Schooling
-0.008 (0.015)
-0.005 (0.023)
Spousal schooling
0.035 (0.015)***
0.016 (0.014)
Own Schooling
-0.046 (0.015)***
-0.054 (0.038)
MTR
Never Smoke
MATR
Exercise
MATR
29
Appendix Table 4 Continued
Alcohol
Consumption
MATR
Spousal schooling
0.507 (0.195)**
-0.017 (0.024)
Own Schooling
-0.232 (0.199)
-0.010 (0.029)
Notes: Estimates are based on the regression: Yij=bo + b1 Own Schoolingij + b2(Spousal Schoolingij + b3Femaleij +
2
(
b4ageij + b5age ij + b6EarlyHealthProblemij + b7 Own Schooling)ij*(Female)ij + b8 ((Spousal Schoolingi)ij*(Female)ji + b9
2
(Age)ij*(Female)ij + b10 (age )ij*(Female)ij + b11 (EarlyHealthProblem)Ij *(Female)ij + uij. Robust standard errors,
clustered by twin pairs in parentheses. ***significant at 1% **significant at 5% *significant at 10%
30
Appendix Table 5: Descriptive Statistics, BFRSS
1987 BFRSS
1994 BFRSS
2000 BFRSS
(1)
(2)
(3)
46.18 (14.83)
47.14 (5.64)
74.98 (3.45)
Background
Characteristics
Age
25-34
0.28 (0.45)
35-44
0.24 (0.42)
45-54
0.15 (0.35)
0.48 (0.50)
55-64
0.16 (0.36)
0.14 (0. 34)
65-75
0.17 (0.37)
0.38 (0.47)
0.58 (0.49)
76-85
0.42 (0.49)
Female
0.56 (0.50)
0.55 (0.50)
0.00 (0.00)
Grades of schooling
12.93 (3.14)
13.46 (2.65)
12.96 (3.04)
Less than 12 grades
0.17 (0.37)
0.09 (0.28)
0.17 (0.37)
12 grades
0.34 (0.47)
0.31 (0.46)
0.30 (0.46)
13-15 grades
0.24 (0.43)
0.28 (0.45)
0.22 (0.41)
16 grades or more
0.25 (0.43)
0.32 (0.47)
0.31 (0.46)
3.78 (1.05)
3.11 (1.12)
26.10 (4.83)
26.35 (3.83)
Self-reported Health
BMI
24.80 (4.26)
Overweight
0.43 (0.49)
0.55 (0.50)
0.63 (0.48)
Obese
0.10 (0.30)
0.17 (0.37)
0.15 (0.35)
Never Smoke
0.45 (0.50)
0.30 (0.46)
Notes: Standard deviations in parentheses.
31
Appendix Table 6: Differential Effect of Schooling in Twin and BFRSS Datasets
Panel A: Combined MATRBFRSS data set
BMI
schooling
twin data
twin data*schooling
N
Overweight
Obese
Never
Smoke
-0.153
-0.015
-0.01
0.017
(0.008)***
(0.001)***
(0.001)***
(0.008)***
1.04
0.125
0.048
0.019
(0.433)**
(0.048)**
-0.033
-0.051
-0.088
-0.011
-0.004
0.002
(0.031)**
(0.004)***
(0.002)*
-0.004
38153
38153
38153
37962
BMI
Overweight
Obese
Panel B: Combined MTRBFRSS data set
Self-reported health
schooling
twin data
twin data*schooling
N
0.1
-0.17
-0.013
-0.011
(0.003)***
(0.012)***
(0.001)***
(0.001)***
1.31
0.301
0.008
-0.002
-0.115
-0.724
-0.081
-0.061
-0.056
-0.029
-0.001
0
(0.008)***
-0.05
-0.005
-0.004
29208
29220
29220
29220
Self-reported health
BMI
Overweight
Obese
Never
Smoke
0.077
-0.081
-0.007
-0.008
0.013
(0.006)***
(0.019)***
(0.002)***
(0.002)***
(0.002)***
0.423
-0.176
0.03
-0.073
-0.184
(0.133)***
-0.451
-0.062
-0.041
(0.060)**
-0.01
-0.028
-0.006
0.001
0.017
-0.01
-0.031
-0.004
-0.003
(0.004)***
6232
6257
6257
6257
6254
Panel C: Combined NASNRC-BFRSS data set
schooling
twin data
twin data*schooling
N
Notes: Estimates are based on the regression: Yi=bo + b1Schoolingi + b2(twin data)i + b3(twin data*Schooling)i +
2
b4agei + b5age i + b6femalei + ui. Twin Data is a dummy variable equal to 1 if the observation is from the twin dataset.
Robust standard errors in parentheses. ***significant at 1% **significant at 5% *significant at 10%
32